JPWO2020044435A1 - Data analysis method, data analysis device, and learning model creation method for data analysis - Google Patents

Abstract

This is a method of analyzing the data to be analyzed by the analysis program 33 using the analysis parameters, in step S2 for creating a plurality of learning parameter sets, and analyzing the plurality of reference data using the plurality of learning parameter sets. Step S4, which executes a programmatic analysis to determine a learning parameter set suitable for analysis, and a reference data group, which is a group of reference data considered to be suitable for analysis, are provided for each of the plurality of learning parameter sets. From step S7 to be associated, step S9 to input unanalyzed data, commonality between the unanalyzed data and each reference measurement data group, and a training parameter set associated with each reference data group, one or more analysis parameters It includes a step S11 of determining an actual analysis parameter set by obtaining each value, and a step S12 of executing an analysis of unanalyzed data by an analysis program using the actual analysis parameter set.

Description

The present invention relates to a technique used when analyzing various data including measurement data obtained by measuring a sample using an analyzer by an analysis program.

A chromatograph mass spectrometer that combines a chromatograph and a mass spectrometer is widely used for identifying and quantifying a target compound contained in a sample. In a chromatographic mass spectrometer, a sample is introduced into a chromatograph column, and multiple substances contained in the sample are separated according to the difference in retention time (RT) and introduced into a mass spectrometer (Mass Spectrometry, MS). The time interval at which the substances separated by the chromatograph are introduced into the mass spectrometer is determined according to the speed of scan measurement (mass scanning speed) in the mass spectrometer. The substance introduced into the mass spectrometer is ionized and then separated and detected according to the mass-to-charge ratio (m / z). As a result, three-dimensional data obtained by plotting the detection intensity of ions with respect to the two axes of retention time (RT) and mass-to-charge ratio (m / z) can be obtained. In this three-dimensional data, the detection intensity (signal intensity) of an ion at each mass-to-charge ratio reflects the content of a substance that produces an ion having that mass-to-charge ratio in the sample.

The total ion current (TIC) can be obtained by integrating the signal strength in the direction of the mass-to-charge ratio (m / z) axis at each point on the retention time (RT) axis of the three-dimensional data. Then, the total ion current chromatogram (TICC) can be obtained by plotting the total ion current along the retention time axis.

If each substance contained in the sample is sufficiently separated from each other by the column of the chromatograph, a monomodal bell-shaped peak appears at the position of the retention time of the substance in the TICC waveform (TICC waveform). By identifying the substance from the mass spectrum at that retention time, it is possible to identify what the substance was eluted during that retention time. A substance is identified by comparing the mass spectrum to be identified with the measured mass spectrum or the theoretical mass spectrum of a known substance stored in a database (Data Base, DB). The comparison items are the mass-to-charge ratio (m / z) value in which the mass peak exists, the intensity of the mass peak, and the like. The degree of concordance (score) of the mass spectrum can be used to quantitatively evaluate how reliable the identification result of the substance is. In addition, the amount of each sample separated by the chromatograph can be estimated from the area and height of the peak on the TICC waveform.

However, if the sample contains a plurality of substances having the same or similar retention times, the plurality of substances will be mixed in the eluate eluted from the chromatograph during the retention time of these substances or the time before and after that. Then, mass peaks derived from a plurality of substances are mixed in the mass spectrum at the holding time and the time before and after the holding time, and peaks derived from a plurality of substances are superimposed on the peaks of the TICC waveform obtained by integrating those mass peaks. It will be the one that was done. Normally, it is considered that some substance was eluted during the retention time of the peak top of the monomodal peak appearing in the TICC waveform, but when it is a superposed peak, the shape of the peak is distorted or a large monomodal peak. Small monomodal peaks are buried in the area, or the peaks are multimodal. In such a case, the retention time at the peak top of the monomodal peak that should be obtained when the eluate from the chromatograph contains only a single substance cannot be correctly determined. Also, if the measurement data contains noise or the signal strength contains baseline components, the situation becomes more complicated and the retention time for small TICC peaks from substances that are only present in small amounts in the sample. Becomes more difficult to find.

Therefore, superposed peaks are separated by signal processing, statistical processing, etc., and peak separation (Peak Deconvolution) is performed to purify the TICC peaks so that one mass spectrum contains only a group of mass peaks derived from a single substance. By purifying the TICC peak in this way, it is possible to estimate what kind of TICC peak was superimposed on the TICC waveform of the measurement data. In many cases, a dedicated analysis program is used to perform peak separation. The National Institute of Standards and Technology is a typical analysis program used to purify the peaks of measurement data (GCMS data) obtained by gas chromatography-mass spectrometry (GC / MS). AMDIS (Automated Mass Spectral Deconvolution and Identification System) provided by NIST) is known (see Non-Patent Document 1). AMDIS uses six analytical parameters (peak width, excluded mass-to-charge ratio, number of neighboring peaks, peak spacing, peak detection sensitivity, and model goodness of fit) to purify the peaks. Initial values are prepared for each of these analysis parameters, and in many cases, the initial values are used as they are for analysis.

Japanese Patent Application Laid-Open No. 2007-41234

"Automated Mass Spectral Deconvolution & Identification System", [online], The National Institute of Standards and Technology (NIST), [Search June 22, 2018], Internet <URL: https://chemdata.nist.gov/ mass-spc / amdis / explanation.html> "Mass ++", [online], Shimadzu Corporation, [Search June 22, 2018], Internet URL: https://www.shimadzu.co.jp/aboutus/ms_r/masspp.html Takayuki Okatani "Deep Learning (Machine Learning Professional Series)" Kodansha, April 2015

The initial values of the analysis parameters in AMDIS are values set assuming general use for various GCMS data, and the state of the GCMS data (shape of superimposed peak, mass scanning speed, noise state, etc.). Some are not always appropriate. That is, even if the initial value is used as it is for the peak separation of the GCMS data to be analyzed, it may not be possible to sufficiently separate the peaks. In such cases, the substance is identified with sufficient reliability, for example, until the user changes the value of each analysis parameter from the initial value to separate the peaks and obtains a result that the analyst considers appropriate. In other words, adjust the parameters until a sufficiently high score is obtained. At this time, since the analyst adjusts the parameters based on the intuition and experience he has cultivated, the analysis result depends on the ability and feeling of the user, that is, the result obtained by the skill level of the analyst. There was a problem that the variation occurred. In addition, there is a problem that the analysis work takes time and effort because it is necessary to repeat the parameter adjustment.

Here, as an example of the prior art, the case of analyzing GCMS data by AMDIS has been described, but various data such as measurement data obtained by measuring a sample using another analyzer can be used with some analysis parameters. There was a similar problem as described above when analyzing.

The problem to be solved by the present invention is to easily obtain appropriate analysis results when analyzing various data such as measurement data obtained by measuring a sample using an analyzer using analysis parameters. It is to provide the technology that can be done.

The first aspect of the present invention, which has been made to solve the above problems, is a method of analyzing data to be analyzed by a predetermined analysis program by setting values for one or a plurality of analysis parameters.
A learning parameter set creation step for creating a plurality of learning parameter sets in which at least one value of at least one of the one or a plurality of analysis parameters is different from each other.
For each of the plurality of reference data, the learning parameter set determination step of executing the analysis by the analysis program using each of the plurality of learning parameter sets and determining the learning parameter set suitable for the analysis according to a predetermined criterion. When,
A reference data group creation step in which a reference data group, which is a group of reference data for which the learning parameter set is determined to be suitable for analysis in the learning parameter set determination step, is associated with each of the plurality of learning parameter sets. ,
Unanalyzed data input step to input unanalyzed data and
The commonality between the unanalyzed data and each reference data group is obtained according to a predetermined standard, and the one or more analysis parameters are obtained from the learning parameter set associated with each reference data group based on the commonality. The actual analysis parameter set determination step, which determines the actual analysis parameter set by obtaining a value suitable for the analysis of unanalyzed data, and the actual analysis parameter set determination step.
It is characterized by including an actual analysis step of executing the analysis of the unanalyzed data by the analysis program using the parameter set for real analysis.

In the data analysis method according to the present invention, first, a plurality of learning parameter sets in which at least one value of one or a plurality of analysis parameters used for data analysis is different from each other are created. Then, each of the plurality of reference data is analyzed by an analysis program using each of the plurality of learning parameter sets, and a learning parameter set suitable for the analysis is determined according to a predetermined standard. For example, for each of the plurality of reference data, an evaluation value indicating the validity of the analysis result obtained by executing the analysis by the analysis program using each of the created plurality of training parameter sets is obtained. This can be done by setting the one with the highest evaluation value as the optimum learning parameter set. Alternatively, a learning parameter set whose evaluation value is equal to or higher than a predetermined reference value may be used as a learning parameter set suitable for analysis. In the former case, one learning parameter set suitable for the analysis is determined for each of the reference data, and in the latter case, one or more learning parameter sets are determined.

Subsequently, for each of the plurality of learning parameter sets, a reference data group which is a group of reference data for which the learning parameter set is considered to be suitable for analysis is created in the learning parameter set determination step. As a result, the reference data having a common learning parameter set suitable for analysis is grouped, and the information that is the basis for determining the parameter set used for the analysis of the unanalyzed data can be obtained. When a plurality of parameter sets suitable for analysis are determined for one reference data in the above training parameter set determination step, the reference data may be included in the plurality of reference data groups.

Next, the unanalyzed data is input. Then, based on the commonality between the unanalyzed data and the reference data group according to a predetermined standard, the parameter set for real analysis is determined by obtaining the values suitable for the analysis of the unanalyzed data for each of the one or more analysis parameters. do. This predetermined criterion differs depending on the type of data to be analyzed. For example, when the data to be analyzed is a TICC waveform of GCMS data, a reference data group composed of reference data having a peak having a shape close to the peak of unanalyzed data is used. , Can be a reference data group with high commonality.

In the actual analysis parameter set determination step, for example, the reference data group having the highest commonality with the unanalyzed data is determined, and the learning parameter set corresponding to the reference data group is used as it is as the actual analysis parameter set. It can be carried out. In this way, when predicting one of the parameter set numbers assigned to a plurality of learning parameter sets, the parameter set number associated with the reference data group having the highest commonality is selected. This is a case where the set of one or more parameters is treated as one "parameter set" and the parameter set to be analyzed should be considered. This is an "identification" approach in machine learning terms.

In addition, as another method of performing the parameter set determination step for actual analysis, the value of each analysis parameter is directly associated with the reference data group (each reference data group may be composed of only one reference data) and unanalyzed. An approach of estimating the value of each analysis parameter to be used in data analysis is also conceivable. This is a "regression" approach in machine learning terms. In the case of regression, unanalyzed data and each reference data group (or each reference) for one analysis parameter, even if the training parameter set contains only two values (eg 5 and 10). Regression analysis is performed based on the commonality with the data) (for example, the similarity of the TICC waveform), and an intermediate value (for example, 7) that is neither of the above two values is the optimum analysis parameter value for the analysis of unanalyzed data. Can be obtained as. Such regression analysis can be performed individually for one or more analysis parameters, or collectively (ie, in parameter set units) for one or more analysis parameters. Finally, the analysis of the unanalyzed data is executed by the analysis program using the parameter set for real analysis composed of the values of one or more analysis parameters obtained by the above regression analysis.

As described above, in the data analysis method according to the present invention, prior to the analysis of the unanalyzed data, the analysis using a plurality of reference data refers to one or a plurality of reference data having a common learning parameter set suitable for the analysis. Group as a data group. Then, based on the commonality between the unanalyzed data and the reference data group, values suitable for analysis of the unanalyzed data for each of the one or more analysis parameters from the learning parameter set associated with the reference data group. And determine them as a parameter set for actual analysis. Therefore, it is not necessary for the user to set the value of the analysis parameter by himself / herself, and the parameter set suitable for the analysis of the unanalyzed data is uniquely determined, so that the appropriate analysis result can be easily obtained. In addition, the results obtained do not vary depending on the skill level of the user.

A second aspect of the present invention made to solve the above problems is used to determine the values of one or more analysis parameters used when analyzing the data to be analyzed by a predetermined analysis program. How to create a learning model
A learning parameter set creation step for creating a plurality of learning parameter sets in which at least one value of at least one of the one or a plurality of analysis parameters is different from each other.
For each of the plurality of reference data, the learning parameter set determination step of executing the analysis by the analysis program using each of the plurality of learning parameter sets and determining the learning parameter set suitable for the analysis according to a predetermined criterion. When,
A reference data group creation step in which a reference data group, which is a group of reference data for which the learning parameter set is determined to be suitable for analysis in the learning parameter set determination step, is associated with each of the plurality of learning parameter sets. ,
It is characterized by having a learning model creation step of creating a learning model by machine learning in which the reference data group is associated with each of the plurality of learning parameter sets as learning data.

In the method for creating a learning model for data analysis, which is the second aspect of the present invention, the same learning parameter creation step, training parameter set determination step, and reference data group creation step as in the data analysis method of the first aspect are performed. A learning model is created by machine learning in which the reference data group is associated with each of a plurality of learning parameter sets created by performing the training data. In recent years, various methods of machine learning have been proposed (for example, Patent Document 1), and the machine learning includes, for example, deep learning, and a convolutional neural network (Convolution Neural Network,) which is a form of the deep learning. CNN), Support Vector Machine (SVM), and AdaBoost can be used. The learning model thus created can be suitably used in the parameter set determination step for actual analysis of the data analysis method according to the first aspect of the present invention.

Further, a third aspect of the present invention made to solve the above problems is an apparatus for analyzing data to be analyzed by a predetermined analysis program by setting values for one or a plurality of analysis parameters. hand,
A learning parameter set creation unit that creates a plurality of learning parameter sets in which at least one value of at least one of the one or a plurality of analysis parameters is different from each other.
A learning parameter set determination unit that executes analysis by the analysis program using each of the plurality of learning parameter sets for each of the plurality of reference data, and determines a learning parameter set suitable for analysis according to a predetermined criterion. When,
A reference data group creation unit that associates each of the plurality of learning parameter sets with a reference data group that is a group of reference data for which the learning parameter set is determined to be suitable for analysis by the learning parameter set determination unit. ,
Unanalyzed data input section for inputting unanalyzed data,
The commonality between the unanalyzed data and each reference data group is obtained according to a predetermined standard, and the one or more analysis parameters are obtained from the learning parameter set associated with each reference data group based on the commonality. A parameter set determination unit for actual analysis that determines a parameter set for actual analysis by obtaining a value suitable for analysis of unanalyzed data, and a parameter set determination unit for actual analysis.
It is characterized by including an actual analysis execution unit that executes analysis of the unanalyzed data by the analysis program using the parameter set for real analysis.

When analyzing various data such as measurement data obtained by measuring a sample using an analyzer using analysis parameters, the data analysis method, data analysis device, or learning model for data analysis according to the present invention. By using the preparation method, an appropriate analysis result can be easily obtained.

FIG. 6 is a configuration diagram of a main part of an analysis system in which a control / processing device, which is an embodiment of the data analysis device according to the present invention, is combined with a gas chromatograph mass spectrometer. The flowchart regarding one Example of the data analysis method which concerns on this invention. An example of a heat map (a) of three-dimensional data obtained by measuring a sample using a gas chromatograph mass spectrometer, and an example of a total ion current chromatogram (b). A description of the analysis parameters used in AMDIS. A part of the learning parameter set used in this embodiment. A histogram showing the analysis result of the divided reference data using a plurality of training parameter sets in this embodiment. For each of the three types of learning parameter sets used in this example, the peaks for which the learning parameter set was optimal for analysis are overlaid. The structure of the data used to evaluate the learning model created by machine learning. The figure explaining the structure of the convolution neutral network used in the machine learning in this Example. The hyperparameters and network configuration of the convolutional neutral network with the highest percentage of correct answers in this example. Correct answer rate when the analysis process of selecting the optimum parameter set according to the learning model of this example is evaluated by 5-fold cross-validation. The figure explaining the process of taking out the division unanalyzed data from unanalyzed data. The conceptual diagram of the data analysis method and analysis apparatus which concerns on this invention. The block diagram of the modification of the data analysis apparatus which concerns on this invention.

Examples of the data analysis method, the data analysis device, and the learning model creation method for data analysis according to the present invention will be described below with reference to the drawings.

The data to be analyzed in this example is three-dimensional GCMS data acquired by measurement using a gas chromatograph mass spectrometer. Further, in this example, AMDIS is used as an analysis program, and the mass spectrum purified by separating the peaks of the waveform (TICC waveform) of the total ion current chromatogram obtained from the GCMS data is obtained from the substance database (substance DB). By collating with the mass spectrum of various known substances stored in advance in the sample, the substances contained in the sample are identified, and the evaluation value (score) indicating the degree of agreement is calculated. This score indicates that the higher the value, the higher the reliability of substance identification.

FIG. 1 is a configuration diagram of a main part of an analysis system including a data analysis device of this embodiment, and FIG. 2 is a flowchart of a data analysis method of this embodiment. The analysis system of this embodiment includes a gas chromatograph mass spectrometer 1 and a control / processing device 3.

The gas chromatograph mass spectrometer 1 includes a gas chromatograph 10 and a mass spectrometer 20. In the gas chromatograph 10, the liquid samples preset in the autosampler 14 are sequentially sent to the injector 13 and injected from the injector 13 into the sample vaporization chamber 12. Further, a carrier gas such as helium is supplied to the sample vaporization chamber 12. The sample vaporization chamber 12 is heated, and the liquid sample injected from the injector 13 is vaporized and rides on the flow of carrier gas, and is sent to the capillary column 15 housed in the column oven 11. Various compounds contained in the sample gas are separated in the time direction while passing through the capillary column 15, and are sequentially introduced into the mass spectrometer 20.

The mass spectrometer 20 includes a vacuum chamber 23 that is evacuated by a vacuum pump (not shown), and an ion source 21, a lens electrode 22, a quadrupole mass filter 24, and an ion detector 25 are arranged therein. ing. The substances in the sample gas introduced from the gas chromatograph 10 are sequentially introduced into the ion source 21. The ion source 21 is, for example, an EI (electron ionization) source, and ions are generated by irradiating the sample gas introduced into the ionization chamber 211 with thermions generated by the filament 212. The ions generated by the ion source 21 are converged by the lens electrode 22 and separated by the quadrupole mass filter 24 according to the mass-to-charge ratio, and then detected by the ion detector 25. The output signal from the ion detector 25 is stored in the storage unit 31 included in the control / processing device 3.

The control / processing device 3 has a function as an analysis control unit that controls each part of the gas chromatograph mass spectrometer 1 and a function of processing data obtained by measurement using the gas chromatograph mass spectrometer 1 or the like. ing. The latter corresponds to the data analysis apparatus according to the present invention. The control / processing device 3 includes a storage unit 31 and a substance database (substance DB) 32, and a predetermined analysis program (AMDIS in this embodiment) 33 is pre-installed. The substance database 32 is a database used for identifying substances contained in a sample in the analysis of data by the analysis program 33, and for each of a large number of known substances, the substance name, chemical formula, theoretical retention time, mass spectrum, etc. Information is associated and saved.

The control / processing device 3 further has, as functional blocks, a reference data acquisition unit 41, a parameter set creation unit 42, a parameter set determination unit 43, a reference data division unit 44, a learning model creation unit 45, and an unanalyzed data input reception unit 46. , Unanalyzed data division unit 47, actual analysis parameter determination unit 48, actual analysis execution unit 49, analysis result output unit 50, and learning model update unit 51. The substance of the control / processing device 3 is a computer, and these functional blocks are embodied by executing the data analysis program 40 pre-installed in the control / processing device 3 on the processor. Further, an input unit 6 such as a mouse or a keyboard and a display unit 7 are connected to the control / processing device 3.

Next, the procedure for analyzing GCMS data in this embodiment will be described together with an actual analysis example with reference to the flowchart of FIG. In addition, steps S1 to S8 in the flowchart of FIG. 2 are procedures of one Embodiment of the learning model creation method which concerns on this invention.

When the user instructs the acquisition of the reference data by the operation through the input unit 6, the reference data acquisition unit 41 operates each unit of the gas chromatograph mass spectrometer 1 and sequentially sets the samples set in the autosampler 14 by the user in order. It is introduced into the gas chromatograph mass spectrometer 1 and each sample is measured. The GCMS data obtained by the measurement of each sample is sequentially stored in the storage unit 31 of the control / processing device 3. Here, the case where the reference data is acquired by actually measuring the sample has been described, but the reference data acquisition unit 41 may read the reference data acquired in advance from the storage unit 31 according to the instruction by the user. good. In this way, a plurality of reference data are acquired (step S1).

In this example, 32 types of biological samples containing a part or all of 504 types of known substances are prepared, and 32 types of GCMS data are measured by performing measurement using a gas chromatograph mass spectrometer 1 for each type. Was acquired. For each of these 504 types of known substances, retention time and mass spectrum information is stored in the substance database 32. The measurement time was between 4 and 24 minutes after sample injection, and within this time, 24000 scans were measured in the mass-to-charge ratio range of 80 to 500.

FIG. 3A shows an example of GCMS data. This is a graph in which the peak intensity of the graph with the retention time (RT) and mass-to-charge ratio (m / z) as the two axes is _{converted to the log 10} scale, and the value is expressed by the difference between cold colors and warm colors (however, the figure). 3 (a) is a monochrome display). In addition, a part of the TICC waveform (data for 40 scans) created from this GCMS data is shown in FIG. 3 (b). In the case of a sample containing a large number of substances as in this example, a plurality of substances having the same or similar retention times are often contained, and the samples are eluted from the chromatograph at the retention time of these substances or the time before and after the retention time. Multiple substances are mixed in the eluate. As a result, mass peaks derived from a plurality of substances are mixed in the mass spectrum at the retention time of those substances and the time before and after that, and the peak of the TICC waveform obtained by integrating those mass peaks is also shown in FIG. As shown in (b), peaks derived from a plurality of substances are superposed (superimposed peaks).

When the user instructs the creation of the learning parameter set by the operation through the input unit 6, the learning parameter set creation unit 42 executes the analysis program 40 installed in the control / processing device 3 in advance and sets the parameters. The screen to be displayed is displayed on the display unit 7. The analysis program of this example is AMDIS. AMDIS uses six analysis parameters: Component width, Omit m / z, Adjacent peak subtraction, Resolution, Sensitivity, and Shape requirement. FIG. 4 shows the contents of each parameter. In this example, 45 types of parameter sets were created by lowering the Adjacent peak subtraction from Two to One and the Resolution from High to Medium based on the initial value (Parameter Set Number: 0). FIG. 5 shows a part (initial value and 10 kinds of parameter sets). Here, the order in which the training parameter set is created after the reference data is acquired has been described, but the execution order of both may be reversed, or both may be performed in parallel. Further, according to the instruction by the user, the learning parameter set creating unit 42 may read the learning parameter set created in advance from the storage unit 31. In this way, a plurality of learning parameter sets are created (step S2).

When the reference data is acquired and the training parameter set is created, the learning parameter set determination unit 43 analyzes each of the 32 GCMS data by AMDIS using 45 types of training parameter sets individually. Is executed (step S3). Specifically, for each set of GCMS data and training parameter set, the peaks of the TICC waveform included in the GCMS data are purified, and the mass spectrum corresponding to each peak is the mass spectrum stored in the substance database 32. The substance corresponding to each peak is identified by matching. Furthermore, the evaluation value (score) is obtained from the degree of coincidence of the mass spectra. AMDIS requires a score of 1 to 100, which indicates the reliability of identification and substance identification. In this example, if the score is 60 or more, the peak is identified as complete, and the peak with a score of less than 60 is unidentified. For the peaks for which identification has been completed, the identified substance name, retention time, parameter set number used for analysis, and score are associated and stored in the storage unit 31.

Next, the learning parameter set determination unit 43 determines the optimum learning parameter set for identification for each of the peaks of the 32 GCMS data (step S4). If the peak with the same retention time is identified by analysis using multiple learning parameter sets, the learning parameter set with the highest score among the multiple learning parameter sets is selected. The optimum training parameter set for peak identification. When there are a plurality of maximum scores, the one with the smaller number of the learning parameter set is set as the optimum learning parameter. Such processing is performed for all peaks (peaks in which substances have been identified) to determine the optimum training parameter set. However, if the difference between the peak retention time and the theoretical retention time of the identified substance is greater than 0.25 minutes, it will be misidentified regardless of the score, and the next largest score and the difference in retention time will be 0.25 minutes or less. A learning parameter set was set as the optimum learning parameter set for identifying the peak.

In this embodiment, when there are a plurality of highest scores, the one with the smaller number of the learning parameter set is the optimum learning parameter set, but the one with the larger number of the learning parameter set is the optimum learning parameter set. Or both may be the optimal learning parameter set.

When the optimum training parameter set for all peaks is determined, the reference data dividing unit 44 extracts (divides the reference data) 40 scans of data centered on the holding time (peak top) for each peak (step S5). ). In this example, 1806 data (divided reference data) were obtained from this. FIG. 6 shows the relationship between these 1806 data and the optimum learning parameter set (histogram showing the number of divided reference data associated with each learning parameter set).

In this embodiment, there are 667 divided reference data in which the initial value of the analysis parameter is the optimum learning parameter set, and the other divided reference data in which the learning parameter set is the optimum analysis parameter set. Was 1,139. As described above, in many cases, there is a certain percentage of data in which the initial values of the analysis parameters are not optimal.

Next, the learning model creation unit 45 selects three learning parameter sets (0, 1, 12) in which the number of associated division reference data is 200 or more from among 45 types of learning parameter sets. Is selected as a candidate for the parameter set for actual analysis to be used for actual analysis, and is extracted together with the divided reference data associated with each. Here, a group of one or a plurality of divided reference data associated with one learning parameter set constitutes one reference data group. That is, the learning model creation unit 45 of this embodiment has a function as a reference data group creation unit according to the present invention. The data created in this way becomes the learning data used in the machine learning described later (step S6). It is possible to extract a training parameter set with less than 200 split reference data associated with it, but if the number of associated split reference data is too small, it is a common feature. It is difficult to identify a specific part (for example, the shape of a peak) by machine learning. Not limited to this embodiment, the number of data required for machine learning analysis differs depending on the type of data to be analyzed and the content of analysis, but in general, it depends on the learning parameter set extracted at this stage. It is desirable that at least several tens to 100 (divided) reference data are associated with each other.

FIG. 7 is an overlay of the TICC waveforms of the divided reference data (40 scans of data standardized with the highest intensity) associated with the learning parameter sets for each of the three learning parameter sets. It is a thing. FIG. 7 (a) is for parameter set 0 (initial value), FIG. 7 (b) is for parameter set 1, and FIG. 7 (c) is for parameter set 12. It is difficult to find the peak shape characteristic of each group (reference data group) only by visually comparing the peak shapes of the TICC waveforms included in FIGS. 7 (a) to 7 (c). .. In addition, many peaks have a peak top appearing in the center, but some peaks do not seem to exist in the center at first glance. By using an analysis program such as AMDIS, it can be seen that peaks that are difficult to extract visually are also extracted.

Next, the learning model creation unit 45 has a total of 1,092 divided reference data (parameter set 0: 667, parameter set 1: 212) associated with each of the three training parameter sets 0, 1 and 12. , Parameter set 12: 213) to create a learning model by machine learning (step S7). In this example, a learning model was constructed using a convolutional neural network (CNN). In addition, the 5-fold cross validation (CV) method was used as a method for evaluating the learning model. The 5-division CV method constructs a learning model using the data of CV numbers 1 to 4, applies it to the data of CV number 0, and calculates the correct answer rate for the data of CV number 0, then CV number 0, The average value of the five correct answer rates obtained by constructing a learning model using the data of 2 to 4 and applying it to the data of CV number 1 to calculate the correct answer rate for the data of CV number 1. Is the performance of the model. In such a cross-validation method, the "data used for model construction" and the "data used for evaluation of the construction model" are different, so it can be said to be a method for evaluating the prediction performance for unknown data. In this embodiment, as shown in FIG. 8, the divided reference data is divided into five data (CV numbers 0 to 4).

The learning model in this embodiment can be regarded as a kind of classifier that outputs a result according to the characteristics of the input data. In this example, CNN is used, but in addition, a learning model can be constructed by using deep learning other than CNN, Support Vector Machine (SVM), AdaBoost, and the like.

FIG. 9 is a schematic configuration diagram of the CNN network used for creating the learning model in this embodiment. In this example, one-dimensional convolution was performed. Then, based on this learning model, the hyperparameters and network configuration (for example, Non-Patent Document 3) having the highest CV correct answer rate were determined (step S8). The result is shown in FIG. As shown in FIG. 11, the average value of the correct answer rate obtained by this hyperparameter and the network configuration was 88.1%. In other words, a prediction model that can predict the optimum parameter set with a probability of about 90% for unknown data (divided unanalyzed data obtained by extracting the part where the peak exists from the GCMS data of the sample whose contained substance is unknown). Was built.

In this embodiment, as described above, 1,092 divided reference data are used as training data. Of these, the initial value (parameter set 0) of the AMDIS analysis parameters was the optimum parameter set for 667 items, that is, 61.1% of the training data. On the other hand, in the learning model created in this example, a correct answer rate of 88.1% was obtained. From these comparisons, by using the learning model of this example, it is more likely than before that the optimum parameter set for data analysis can be selected and the substances contained in the sample can be identified with the highest accuracy. It can be said that it was.

When the learning model is created by the learning model creating unit 45, the unanalyzed data input receiving unit 46 displays a screen for inputting the data to be analyzed on the display unit 7. The user measures the sample set in the autosampler 14 with the gas chromatograph mass spectrometer 1 and inputs the acquired GCMS data as unanalyzed data, as in the case of acquiring the reference data. Alternatively, when analyzing the already measured data, the unanalyzed data previously stored in the storage unit 31 is read out and input. The unanalyzed data of this example is GCMS data obtained by 24000 scans of the mass-to-charge ratio range of 80 to 500 within 4 to 24 minutes after sample injection, similar to the reference data. In this way, the data to be analyzed is input as unanalyzed data (step S9). The measurement conditions for the unanalyzed data do not necessarily have to be the same as the measurement conditions for the reference data.

When the unanalyzed data is input, the unanalyzed data dividing unit 47, as shown in FIG. 12, 40 scans while shifting the extraction start position, for example, by 10 scans from the side where the holding time of the input unanalyzed data is short. Data is taken out. As a result, 2397 divided unanalyzed data are created from the unanalyzed data (step S10).

When the divided unanalyzed data is obtained, the actual analysis parameter determination unit 48 inputs the divided unanalyzed data one by one into the training model as unknown data, and the divided unanalyzed data from the parameter sets 0, 1, 12 is input. Output the most suitable parameter set for data analysis. The training model determines a reference data group having the highest commonality with the peak characteristics included in the divided unanalyzed data, and determines an actual analysis parameter set corresponding to the reference data group (step S11).

Normally, not all divided unanalyzed data generated from one unanalyzed data contains peaks, and only a part of them has peaks. For the divided unanalyzed data that does not include peaks, there is no reference data group that has the same characteristics as the data, so there is no optimal parameter set. Therefore, it is determined that there is no analysis target (peak) for such divided unanalyzed data, and the optimum parameter set is selected only for the divided unanalyzed data in which the analysis target (peak) exists.

The real analysis execution unit 49 performs analysis by AMDIS using the parameter set selected by the learning model to purify the peaks, identifies the substance corresponding to each peak, and obtains the score (step S12). When the real analysis execution unit 49 completes the identification of the substance corresponding to the peak, the analysis result output unit 50 displays (outputs) the name, retention time, and score of the identified substance on the display unit 7 (step S13). .. In addition to this information, the number of the parameter set used for peak identification may be output. Further, in step S12, when the retention time of the identified peak and the theoretical retention time of the identified substance differ by a predetermined time (for example, 0.25 minutes) or more, the identification result is discarded (or a warning is added. ) Can be added. As a result, the possibility of misidentification due to accidental matching of mass spectra can be eliminated, and the identification accuracy can be improved.

The main purpose of the data analysis method and the apparatus in this embodiment is to analyze the unanalyzed data described above, but the data analysis apparatus of the present embodiment further includes a learning model update unit 51.

When a predetermined number (for example, 30) of unanalyzed data analyzed by the real analysis execution unit 49 (hereinafter, this is referred to as “analyzed data”) is accumulated, the learning model update unit 51 has analyzed them. Set the data to the reference data described above. When the reference data is set in this way, the same processing as in steps S1 to S8 described above is performed in order. Then, the hyperparameters and network configuration of the learning model are readjusted so that the correct answer rate in the 5-division CV method is the highest, and the learning model is updated. In this way, by sequentially using the analyzed data as reference data, the learning model can be updated so as to be able to handle a wider variety of data. Here, the learning model is updated by the learning model update unit 51 every time a predetermined number of analyzed data is accumulated, but the learning model is updated (reconstructed) every time the analyzed data is generated. You may. Here, the case of updating the learning model by online learning (sequential learning) that executes machine learning using only the newly added analyzed data as reference data has been described as an example, but the reference data and analysis already used for machine learning have been described. The learning model may be updated by batch learning using both the completed data.

FIG. 13 schematically shows the concept of the data analysis method and the analysis device of this embodiment. As shown in FIG. 13, in this embodiment, a learning model (prediction model f (x)) as a discriminator that outputs a result f (x) according to the characteristics of the input data x is prepared in advance by machine learning. Create it. When the GCMS data to be analyzed is input, undivided unanalyzed data is created from the GCMS data, and the waveform data of the total ion current chromatogram created from the unanalyzed data is input to the training model to obtain the optimum parameter set. It is output. Then, using this as an analysis parameter, peak purification (peak separation, etc.) by AMDIS and identification of the substance corresponding to the peak are performed, and the results (identification substance name and identification score) are output.

In the data analysis method and data analysis device of this embodiment, one of the training parameter sets is set as the most suitable parameter set (actual analysis parameter set) for the data to be analyzed by the learning model created by machine learning. It is selected and analyzed by AMDIS using the parameter set for actual analysis. Therefore, the user does not have to change the value of the analysis parameter by himself / herself, and the optimum analysis result can be easily obtained with a high probability. In addition, there is no difference in the analysis result depending on the skill level of the user. Furthermore, since the learning model is updated every time a predetermined number of analyzed data are accumulated, it is possible to always analyze various data with high accuracy.

Next, a modified example of the data analysis device according to the present invention will be described. In the data analysis device (control / processing device 3) of the above embodiment, both the learning model was created and the data was analyzed, but in the data analysis device of the modified example, the data was generated using the learning model created in advance. To analyze.

FIG. 14 is a block diagram of a control / processing device 3a which is a modification of the data analysis device according to the present invention. The components common to the control / processing device 3 of the above embodiment are designated by the same reference numerals, and the description thereof will be omitted as appropriate. Similar to the above embodiment, the substance of the control / processing device 3a of the modified example is also a personal computer, and each functional block shown in FIG. 14 is embodied by executing the analysis program 40a.

A learning model (CNN) 34 corresponding to an analysis program (AMDIS in the above embodiment) is pre-installed in this control / processing device 3a, and a learning parameter set used when constructing the learning model 34 is installed. It differs from the control / processing device 3 of the above embodiment in that it is stored in the storage unit 31a. This learning model 34 is a port of the learning model 34 created by executing steps S1 to S8 described in the above embodiment, and is a shipment of a personal computer configured as a control / processing device 3a of a modified example. Installed at the previous stage.

Therefore, the user of the control / processing device 3a of the modified example can analyze the data using the learning model by executing only steps S9 to S13 without executing S1 to S8 of the above embodiment by himself / herself. can.

The control / processing device 3a of the modified example also includes the learning model update unit 51a as in the above embodiment, and like the above embodiment, the learning model update unit 50a is used every time a predetermined number of analyzed data are accumulated. The parameters and network configuration of the learning model 34 are updated as appropriate. In the above embodiment, the learning model is updated by either batch learning or online learning, but in the control / processing device 3a of the modified example, the learning model is updated by online learning.

The above embodiment is an example and can be appropriately modified according to the gist of the present invention.
In the above embodiment, the learning parameter set determination unit 43 determines the optimum learning parameter set for each peak, but the learning parameter for which a score (evaluation value) equal to or higher than a predetermined value is obtained. The entire set may be a parameter set suitable for analysis. Alternatively, all the training parameter sets obtained with a score of a certain ratio (for example, 90%) or more with respect to the highest score obtained for the peak of the same retention time can be set as a parameter set suitable for analysis. In these cases, the same peak data (division reference data) is associated with a plurality of analysis parameters.

Further, in the above embodiment, the divided unanalyzed data is created from the entire unanalyzed data and input to the training model, but the portion where the peak (analysis target) exists is extracted in advance from the unanalyzed data. The divided unanalyzed data may be created only from that part. For example, the initial values of the analysis parameters are used as they are, and the unanalyzed data is analyzed by AMDIS to extract the peaks, or another peak detection software is used to identify the part where the peaks exist from the unanalyzed data. You may try to do it. Furthermore, the user may specify the range in which the peak is considered to exist.

Further, in the above embodiment, a set of values of one or more analysis parameters is set as one learning parameter set, and the most suitable one among the plurality of learning parameters for analysis of unanalyzed data is the actual analysis parameter set. And said. That is, the case of predicting one of the categories of parameter set numbers corresponding to a plurality of training parameter sets prepared in advance has been described as an example. This is an "identification" approach in machine learning terms.

On the other hand, the parameter set for real analysis can be determined by the "regression" approach. Specifically, the value of each analysis parameter is directly associated with the reference data group (each reference data group may be composed of only one reference data), and the value of each analysis parameter to be used in the analysis of unanalyzed data. Is the approach of directly estimating. This approach involves unanalyzed data and each reference data group (or each reference data) for a single analysis parameter, even if the training parameter set contains only two values (eg 5 and 10). ) Is performed, and an intermediate value (for example, 7) that is neither of the above two values is used as the optimum analysis parameter value for the analysis of unanalyzed data. Can be sought. Such regression analysis can be performed individually for each of one or more analysis parameters, or collectively (ie, in parameter set units) for one or more analysis parameters.

In the above example, the case of identifying the substance contained in the sample by using the three-dimensional data obtained by the measurement of the sample using the gas chromatograph mass analyzer as data has been described, but the data analysis method and data analysis according to the present invention have been described. The device and the method of creating a learning model can be widely used for analyzing various data.

For example, Mass ++ is one of the software for analyzing mass spectrometric data of a sample (see Non-Patent Document 2). Mass ++ reads LCMS data obtained by measuring samples containing peptides and proteins with a liquid chromatograph mass spectrometer (including MALDI), smoothing chromatograms and mass spectra, removing baselines, detecting peaks, etc. It is possible to perform analysis by creating a peak list of mass spectrum and sending it to a database search server (Mascot server) to identify peptides and identify proteins predicted from the identified peptides. It is software. Similar to AMDIS, Mass ++ also requires a score (reliability score) that indicates the reliability of identification of identified peptides and proteins.

Various analysis parameters are used when creating a list of mass spectra from LCMS data using Mass ++. In addition, various analysis parameters are also used to identify the substance corresponding to the created peak list. Conventionally, it has been necessary to perform analysis using the initial values of these analysis parameters as they are, or to change the analysis parameters based on the user's own experience. However, by applying the present invention, it is easily optimized. Identification results can be obtained.

In addition, the sample is obtained by analyzing the spectral spectrum data obtained by measuring the sample using an analyzer other than a chromatograph or a mass spectrometer, for example, a spectroscopic measurement device such as a Fourier-converted infrared spectrophotometer, using a predetermined analysis program. The present invention can also be applied to analysis such as identification of a substance contained in. Furthermore, it can also be used for analysis of data obtained by a nuclear magnetic resonance apparatus (NMR), a near-infrared optical brain function imaging apparatus (NIRS), or the like. Furthermore, it can be used for analysis such as predicting future stock price fluctuation data from the latest stock price fluctuation data based on past stock price fluctuation data. That is, the evaluation value becomes high by performing the analysis using the optimum analysis parameters (for example, the correct answer rate for a given question becomes high, the purity of the target substance increases, the power consumption decreases, and the profit increases. The present invention can be applied to various data analyzes as long as it can be defined.

In the above embodiment, comprehensive analysis using a plurality of training parameter sets is performed for all the reference data (and analyzed data), and only the training data for which the optimum parameter set is known in advance for all peaks. A learning model was created by performing so-called supervised learning using Can also be created.

In the above examples and modifications, the case of using batch learning and online learning as machine learning methods has been described, but in addition, transfer learning (learning model using learning data belonging to a domain different from the domain in which the learning model was created). In exchange for the absence of a teacher who explicitly indicates the output for the input, the reward is given as an evaluation of the quality of the result for a series of actions, and the action and the result are Various methods can be used, such as learning the behavior that maximizes the reward by trial and error while updating the information).

For example, the manufacturer of the personal computer, which is the control / processing device 3a of the modified example, has installed CNN34 for judging the quality of the cells of the clone cultured at the company, and the person who purchased this has installed the clone. In addition to determining the quality of cells, it may be used to determine the quality of cells in undifferentiated maintenance culture. That is, when the CNN34 is used for analysis of data acquired in an environment different from the environment in which it was created, the learning model update unit 51a acquires the CNN34 created for determining the quality of the cloned cells in another environment. It will be updated according to the data. Even when such transfer learning is performed, it is possible to use the configurations described in the above examples and modifications. In addition, a learning model created to remove noise from the image data obtained by imaging the sample with an optical microscope and detect the characteristic structure of the sample is obtained by mass spectrometry of the sample using an imaging mass spectrometer. Transfer learning is also performed when applying it to analysis such as removing noise from the acquired data and detecting the characteristic structure of the sample.

Also, for the behavior of changing control parameters such as voltage and temperature of the mass spectrometer, learn the behavior of adjusting the control parameters so as to maximize the reward, using the intensity of the peak obtained as a result of the measurement as a reward. In such cases, the method and apparatus according to the present invention can be used.

1 ... Gas chromatograph mass spectrometer 10 ... Gas chromatograph unit 11 ... Column oven 12 ... Sample vaporization chamber 13 ... Injector 14 ... Auto sampler 15 ... Capillary column 20 ... Mass spectrometer 21 ... Ion source 211 ... Ionization chamber 212 ... Filament 22 ... Lens electrode 23 ... Quadrupole mass filter 23 ... Vacuum chamber 24 ... Ion detectors 3, 3a ... Control / processing devices 31, 31a ... Storage unit 32 ... Material database 33 ... Analysis program 34 ... CNN
40, 40a ... Data analysis program 41 ... Reference data acquisition unit 42 ... Learning parameter set creation unit 43 ... Learning parameter set determination unit 44 ... Reference data division unit 45 ... Learning model creation unit 46 ... Unanalyzed data input reception unit 47 ... Unanalyzed data division unit 48 ... Actual analysis parameter set determination unit 49 ... Actual analysis execution unit 50 ... Analysis result output unit 51, 51a ... Learning model update unit 6 ... Input unit 7 ... Display unit

A chromatograph mass spectrometer that combines a chromatograph and a mass spectrometer is widely used for identifying and quantifying a target compound contained in a sample. In the chromatograph mass spectrometer, a sample is introduced into a chromatograph column, and a plurality of substances contained in the sample are separated according to the difference in retention time (RT) and introduced into a mass spectrometer (Mass Spectrometer , MS) . After substances are introduced into the mass spectrometer that is ionized and detected is separated according to mass-to-charge ratio (m / z). As a result, three-dimensional data obtained by plotting the detection intensity of ions with respect to the two axes of retention time (RT) and mass-to-charge ratio (m / z) can be obtained. In this three-dimensional data, the detection intensity (signal intensity) of ions at each mass-to-charge ratio reflects the content of the substance that produces ions having that mass-to-charge ratio in the sample.

The total ion current (TIC) can be obtained by integrating the signal strength in the direction of the mass-to-charge ratio (m / z) axis at each point on the retention time (RT) axis of the three-dimensional data. The total ion current chromatogram (Total Ion Current Chromatogram, TICC) is obtained by plotting along the retention time axis total ion current.

If each substance contained in the sample is sufficiently separated from each other by the column of the chromatograph, a monomodal bell-shaped peak appears at the position of the retention time of the substance in the TICC waveform (TICC waveform). By identifying the substance from the mass spectrum at that retention time, it is possible to identify what the substance was eluted during that retention time. A substance is identified by comparing the mass spectrum to be identified with the measured mass spectrum or the theoretical mass spectrum of a known substance stored in a database (Data Base, DB). The comparison item, mass-to-charge ratio mass peaks are present (m / z) value, the intensity of Ma-speak and the like. The degree of concordance (score) of the mass spectrum can be used to quantitatively evaluate how reliable the identification result of the substance is. Also, from the area or height of the peak of TICC waveform, it is possible to estimate the amount of each material separated by chromatography.

Therefore, superposed peaks are separated by signal processing, statistical processing, etc., and peak separation (Peak Deconvolution) is performed to purify the TICC peaks so that one mass spectrum contains only a group of mass peaks derived from a single substance. By purifying the TICC peak in this way, it is possible to estimate what kind of TICC peak was superimposed on the TICC waveform of the measurement data. In many cases, a dedicated analysis program is used to perform peak separation. Typical analysis program is used to purify the peak of gas chromatography / mass spectrometry (GC / MS) measurement data obtained in (GCMS data), National Institute of Standards and Technology (National Institute of Standards and Technology AMDIS (Automated Mass Spectral Deconvolution and Identification System) provided by NIST) is known (see Non-Patent Document 1). AMDIS uses six analytical parameters (peak width, excluded mass-to-charge ratio, number of neighboring peaks, peak spacing, peak detection sensitivity, and model goodness of fit) to purify the peaks. Initial values are prepared for each of these analysis parameters, and in many cases, the initial values are used as they are for analysis.

In the actual analysis parameter set determination step, for example, the reference data group having the highest commonality with the unanalyzed data is determined, and the learning parameter set corresponding to the reference data group is used as it is as the actual analysis parameter set. Can be done . When predicting the one of attached to learning parameter set multiple parameter set number, selects a parameter set number most common property is associated with the high reference data group. This is a case where the set of one or more parameters is treated as one "parameter set" and the parameter set to be analyzed should be considered. This is an "identification" approach in machine learning terms.

Further, as functional blocks, the control / processing device 3 includes a reference data acquisition unit 41, a learning parameter set creation unit 42, a learning parameter set determination unit 43, a reference data division unit 44, a learning model creation unit 45, and unanalyzed data. It includes an input reception unit 46, an unanalyzed data division unit 47, an actual analysis parameter determination unit 48, an actual analysis execution unit 49, an analysis result output unit 50, and a learning model update unit 51. The substance of the control / processing device 3 is a computer, and these functional blocks are embodied by executing the data analysis program 40 pre-installed in the control / processing device 3 on the processor. Further, an input unit 6 such as a mouse or a keyboard and a display unit 7 are connected to the control / processing device 3.

When the user instructs the creation of the learning parameter set by the operation through the input unit 6, the learning parameter set creation unit 42 executes the analysis program 33 installed in the control / processing device 3 in advance and sets the parameters. The screen to be displayed is displayed on the display unit 7. The analysis program of this example is AMDIS. AMDIS uses six analysis parameters: Component width, Omit m / z, Adjacent peak subtraction, Resolution, Sensitivity, and Shape requirement. FIG. 4 shows the contents of each parameter. In this example, 45 types of parameter sets were created by lowering the Adjacent peak subtraction from Two to One and the Resolution from High to Medium based on the initial value (Parameter Set Number: 0). FIG. 5 shows a part (initial value and 10 kinds of parameter sets). Here, the order in which the training parameter set is created after the reference data is acquired has been described, but the execution order of both may be reversed, or both may be performed in parallel. Further, according to the instruction by the user, the learning parameter set creating unit 42 may read the learning parameter set created in advance from the storage unit 31. In this way, a plurality of learning parameter sets are created (step S2).

However, if the sample contains a plurality of substances having the same or similar retention times, the plurality of substances will be mixed in the eluate eluted from the chromatograph during the retention time of these substances or the time before and after that. Then, mass peaks derived from a plurality of substances are mixed in the mass spectrum at the holding time and the time before and after the holding time, and peaks derived from a plurality of substances are superimposed on the peaks of the TICC waveform obtained by integrating those mass peaks. It will be the one that was done. Normally, it is considered that some substance was eluted during the retention time of the peak top of the monomodal peak appearing in the TICC waveform, but when it is a superposed peak, the shape of the peak is distorted or a large monomodal peak. Small monomodal peaks are buried in the area, or the peaks are multimodal. In such a case, it can not be determined correctly the retention time corresponding to pin Kutoppu. Also, if the measurement data contains noise or the signal strength contains baseline components, the situation becomes more complicated and the retention time for small TICC peaks from substances that are only present in small amounts in the sample. Becomes more difficult to find.

The second in a is data analysis learning model creation method aspects, the first aspect of the data analysis similar parameter set creation step for learning and methods, learning parameter set determining step, and the reference data group creation step of the present invention A learning model is created by machine learning in which the reference data group is associated with each of a plurality of learning parameter sets created by performing the above as training data. In recent years, various methods of machine learning have been proposed (for example, Patent Document 1), and the machine learning includes, for example, deep learning, and a convolutional neural network (Convolution Neural Network,) which is a form of the deep learning. CNN), Support Vector Machine (SVM), and AdaBoost can be used. The learning model thus created can be suitably used in the parameter set determination step for actual analysis of the data analysis method according to the first aspect of the present invention.

Next, the learning parameter set determination unit 43 determines the optimum learning parameter set for identification for each of the peaks of the 32 GCMS data (step S4). If the peak with the same retention time is identified by analysis using multiple learning parameter sets, the learning parameter set with the highest score among the multiple learning parameter sets is selected. The optimum training parameter set for peak identification. When there are a plurality of maximum scores, the one with the smaller number of the learning parameter set is set as the optimum learning parameter set. Such processing is performed for all peaks (peaks in which substances have been identified) to determine the optimum training parameter set. However, if the difference between the peak retention time and the theoretical retention time of the identified substance is greater than 0.25 minutes, it will be misidentified regardless of the score, and the next largest score and the difference in retention time will be 0.25 minutes or less. A learning parameter set was set as the optimum learning parameter set for identifying the peak.

In this embodiment , there are 667 divided reference data in which the learning parameter set 0, which is the initial value of the analysis parameters, is the optimum learning parameter set, and the other learning parameter sets are the optimum learning parameter sets. The number of divided reference data determined to be is 1,139. As described above, in many cases, there is a certain percentage of data in which the initial values of the analysis parameters are not optimal.

Claims (14)

This is a method of analyzing the data to be analyzed by setting values for one or more analysis parameters and using a predetermined analysis program.
A learning parameter set creation step for creating a plurality of learning parameter sets in which at least one value of at least one of the one or a plurality of analysis parameters is different from each other.
For each of the plurality of reference data, the learning parameter set determination step of executing the analysis by the analysis program using each of the plurality of learning parameter sets and determining the learning parameter set suitable for the analysis according to a predetermined criterion. When,
A reference data group creation step in which a reference data group, which is a group of reference data for which the learning parameter set is determined to be suitable for analysis in the learning parameter set determination step, is associated with each of the plurality of learning parameter sets. ,
An unanalyzed data input step for inputting unanalyzed data, which is unanalyzed data,
The commonality between the unanalyzed data and each reference data group is obtained according to a predetermined standard, and the one or more analysis parameters are obtained from the learning parameter set associated with each reference data group based on the commonality. The actual analysis parameter set determination step, which determines the actual analysis parameter set by obtaining a value suitable for the analysis of unanalyzed data, and the actual analysis parameter set determination step.
A data analysis method comprising: an actual analysis step of executing an analysis of the unanalyzed data by the analysis program using the parameter set for actual analysis. Moreover,
It has a learning model creation step of creating a learning model by machine learning in which the reference data group is associated with each of the plurality of learning parameter sets as learning data.
The data analysis method according to claim 1, wherein a parameter set is determined using the learning model in the parameter selection step. The data analysis method according to claim 2, wherein the machine learning uses deep learning, a support vector machine, and AdaBoost. Moreover,
By executing the learning parameter set determination step using the unanalyzed data as the reference data, a learning parameter set suitable for the analysis is determined, and the unanalyzed data corresponds to the learning parameter set suitable for the analysis. The data analysis method according to claim 2, further comprising a learning model update step for performing the machine learning using the attached data as training data. The data analysis method according to claim 1, wherein the reference data and the unanalyzed data are mass chromatograms, total ion current chromatograms, mass spectra, spectroscopic spectra, or image data. In the parameter set determination step, a parameter set suitable for the analysis is determined for a part or all of the divided reference data obtained by dividing the reference data.
The data analysis method according to claim 1, wherein the reference data group is created by grouping the divided reference data in the reference data group creation step. The analysis program extracts data of one or more peaks included in the unanalyzed data and collates it with a database of known substances to identify substances corresponding to the one or more peaks. The data analysis method according to claim 6. The data analysis according to claim 7, wherein the degree of agreement between the identified substance and the data stored in the database is obtained for each of the data of one or a plurality of peaks included in the unanalyzed data. Method. The data analysis method according to claim 8, wherein the predetermined criterion in the optimum parameter determination step is the one having the highest degree of agreement as the optimum learning parameter set. Multiple divided unanalyzed data are created by dividing the data obtained by measuring the sample to be analyzed using an analyzer according to a predetermined standard.
The data analysis method according to claim 6, wherein a part or all of the plurality of divided unanalyzed data is input as the unanalyzed data in the unanalyzed data input step. The measurement data analysis method according to claim 10, wherein the divided unanalyzed data is data of one or a plurality of peaks. The first aspect of the present invention is to determine the actual analysis parameter set only when there is a reference data group having a high degree of commonality equal to or higher than a predetermined reference in the actual analysis parameter set determination step. The described measurement data analysis method. A device that analyzes the data to be analyzed by setting values for one or more analysis parameters and using a predetermined analysis program.
A learning parameter set creation unit that creates a plurality of learning parameter sets in which at least one value of at least one of the one or a plurality of analysis parameters is different from each other.
A learning parameter set determination unit that executes analysis by the analysis program using each of the plurality of learning parameter sets for each of the plurality of reference data, and determines a learning parameter set suitable for analysis according to a predetermined criterion. When,
A reference data group creation unit that associates each of the plurality of learning parameter sets with a reference data group that is a group of reference data for which the learning parameter set is determined to be suitable for analysis by the learning parameter set determination unit. ,
Unanalyzed data input section for inputting unanalyzed measurement data,
The commonality between the unanalyzed data and each reference data group is obtained according to a predetermined standard, and the one or more analysis parameters are obtained from the learning parameter set associated with each reference data group based on the commonality. A parameter set determination unit for actual analysis that determines a parameter set for actual analysis by obtaining a value suitable for analysis of unanalyzed data, and a parameter set determination unit for actual analysis.
A measurement data analysis apparatus including an actual analysis execution unit that executes analysis of the unanalyzed data by the analysis program using the actual analysis parameter set. A method of creating a learning model used to determine the values of one or more analysis parameters used when analyzing the data to be analyzed by a predetermined analysis program.
A learning parameter set creation step for creating a plurality of learning parameter sets in which at least one value of at least one of the one or a plurality of analysis parameters is different from each other.
For each of the plurality of reference data, the learning parameter set determination step of executing the analysis by the analysis program using each of the plurality of learning parameter sets and determining the learning parameter set suitable for the analysis according to a predetermined criterion. When,
A reference data group creation step in which a reference data group, which is a group of reference data for which the learning parameter set is determined to be suitable for analysis in the learning parameter set determination step, is associated with each of the plurality of learning parameter sets. ,
A learning model for data analysis, which comprises a learning model creation step for creating a learning model by machine learning, in which each of the plurality of learning parameter sets is associated with the reference data group as learning data. How to make.

Images